<h2 id="intro-kyp4u52a1">Intro</h2>This is a lora lab following the previous Lab 1 =&gt; 2 =&gt; 3 =&gt; 4 ( in which I was trying to fix the "potentially black pantyhose" thing )However, my main question of this series is HOW COULD I KEEP ENHANCING THE QUALITY &amp; EFFECT OF CERTAIN LORA ?By distilling from my previous LoRA, now lab_10_4 is kinda acceptable for now ( but still, it's not fully tested yet. I will keep testing around, but play and enjoy if you like )<h3 id="which-base-model-z42ys8trg">WHICH BASE MODEL ?</h3><ul><li>REMEMBER to switch to the correct base model<ul><li>for 10-04, 10-10, 10-11 :<ul><li>Janku v3 : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1277670/janku-v30-noobai-eps-rouwei-nsfw-illustrious-xl">https://civarchive.com/models/1277670/janku-v30-noobai-eps-rouwei-nsfw-illustrious-xl</a></li></ul></li></ul></li></ul><h2 id="-ysvm37ruu">更新</h2><ul><li>overall<ul><li>lora 文档 (先中文 =&gt; 后英文, 可按需阅读)</li><li>lora documentation (Chinese first =&gt; English second, read as needed)</li></ul></li><li>250425<ul><li>上传更新 : OK 10_11, BAD 10_10, fixed dataset &amp; regset</li><li>Upload update: OK 10_11, BAD 10_10, fixed dataset &amp; regset</li><li>提供训练的思路</li><li>Provide training ideas</li></ul></li></ul><h2 id="-mblir5y86">如何使用</h2>简而言之：<ul><li>使用 beige pantyhose<ul><li>如果感觉没有"肤色感"，可加上较弱的 [[yellow pantyhose]]（权重可以再调整）</li><li>use <code>beige pantyhose</code> and maybe somehow add [[yellow pantyhose]] can make it "a bit naturally nude" ( weight should you try around )</li></ul></li><li>lora 权重，感觉 0.5~0.6 就差不多了</li><li>for lora weight, maybe 0.5 ~ 0.6 ( the dataset is to be fixed on next round to make the color more natural )</li><li>使用 sheer golden pantyhose, 而非 单独 golden pantyhose ( golden pantyhose 的数据图片不是太好, 颜色过黄而不自然 从 benchmark 中也可以看出 )</li><li><code>sheer golden pantyhose</code> might be used, but never use <code>golden pantyhose</code> alone! ( it's too strong, as I figure out that the dataset is not that good ... you can see from the benchmark or dataset uploaded )</li></ul><h2 id="-5mvxm7vq3">优化与发现</h2>首先, 展示一下 04 =&gt; 10 =&gt; 11 的 benchmark 的其中一个对比.First, let's see the improvements from 04 to 10 to 11.<ul><li>04 : with weight increase, the image's structure change significantly and a lot of pitfalls can be pinpointed</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/4e7baa23-b057-4a65-8d64-c8204401a55a/width=525/4e7baa23-b057-4a65-8d64-c8204401a55a.jpeg" /><ul><li>10 : basic structure keeps ok... but still, the bad reg set kinda mess the image up</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f46f905f-ab59-40c6-9fa6-85e2bbcad591/width=525/f46f905f-ab59-40c6-9fa6-85e2bbcad591.jpeg" /><ul><li>11 : As it shows, although the color becomes strange ( due to the dataset's color problem ), the image remains stable. There's fewer nonsense problem can be found.</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/080a950f-9354-4b40-b2bb-adbb2411617e/width=525/080a950f-9354-4b40-b2bb-adbb2411617e.jpeg" />下面列举本次 lora 训练的改进方式，方便有兴趣训练 lora 的人参考。<h3 id="1)-()-jqlmy9hxu">1) (非常重要) 图片质量问题可能存在于细节</h3>下面是两个对比例子<ul><li>example 1 : 仔细看整体 &amp; 左侧(有问题) 衣物像素的抖动</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/dc03ed7c-f175-4c02-b912-5ee8c1d13de4/width=525/dc03ed7c-f175-4c02-b912-5ee8c1d13de4.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6ba17aba-ec82-4dce-aa2d-9e50dfbc7a00/width=525/6ba17aba-ec82-4dce-aa2d-9e50dfbc7a00.jpeg" /><ul><li>example 2 : 丝袜区域放大后, 左侧 (有问题) 的颜色梯度变化</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8b1926ff-696a-43df-8c03-fd67f04fab74/width=525/8b1926ff-696a-43df-8c03-fd67f04fab74.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d03e8f6e-b87a-4ac8-b29e-3af916bd2bd4/width=525/d03e8f6e-b87a-4ac8-b29e-3af916bd2bd4.jpeg" />可以看到，从整体画面是没什么大问题，但由于我为了提升图片质量 (下述的 2) )，之前进行的方式是：<ul><li>先生成原始图片</li><li>再 img2img &amp; adetailer 放大 1.5 倍</li><li>再用程序缩小回训练的尺寸 &lt;=== 这一步出的问题</li></ul>检查图片质量时，我没有放大到像素级别，否则我会看到很多的噪点和颜色抖动。人感知不到，而 sd 模型是能充分感知到的。我将 3 次的一些对比实验也给出：<ul><li>10_11：fixed dataset &amp; reg set</li><li>10_10：only fixed dataset &amp; wrong reg set</li><li>10_04：wrong dataset &amp; reg set</li></ul>你可以看到，10_04 在权重提升时，图片构图变化巨大，且这种变化是毫无意义的...如果你恰好也用程序 (nodejs sharp 依赖) 进行 png 处理, 那么可参考如下:<pre><code>const sharpInstance = sharp(inputPath)
 .resize(
 width,
 height,
 {kernel: sharp.kernel.lanczos3}
 )
 .png({
 quality: 100,
 // 0~9 is all lossless compression, 9 is better
 compressionLevel: 9,
 // @note. ****** very very very important ******
 palette: false
 });
</code></pre><h3 id="2)-ciyvlb20j">2) 图片质量的一个技巧处理</h3>虽然 2) 是导致 1) 出现的原因，但我还是想推荐它：一开始我获取一些用于 reg 的图片，是直接用训练的尺寸，生成一批图，然后从里面挑选。但后来发现，先 highres fix，或者用 img2img 放大，图片在宏观上的质量会有明显提升，尤其在人物较小的情况下，如下图 (左侧是未经放大再缩小的) ：<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/bbc83c39-6434-49c6-ab5e-6a500883cc0e/width=525/bbc83c39-6434-49c6-ab5e-6a500883cc0e.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/994a6214-de31-4ad1-8798-378f05016e99/width=525/994a6214-de31-4ad1-8798-378f05016e99.jpeg" />所以，这个处理思路就是：<ul><li>先根据 prompt 生成一大批图片，再选择一批合理且符合预期的图片</li><li>img2img &amp; adetailer 放大</li><li>再缩小到训练尺寸</li></ul><h3 id="3)-reg-set-i3t5folik">3) reg set 的管理</h3>reg 的收集也和 dataset 一样花时间，而且单次训练中，它的数量需求更大，有可能花费更多时间。但好消息是，大多数质量好的 reg set，可以在不同的训练里重复复用 🤩；可惜又有个坏消息是，为了这种重复利用，管理它们也是个头大的事情。😂经验证发现：（至少在 mac &amp; linux 环境下）软链方式下，图片等资源，也能正常作为资源文件被训练脚本使用。所以推荐以下方式：<ul><li>准备一个文件夹（结构大致如下）：<ul><li>my_reg_set/<ul><li>all_images/<ul><li>1girl/ ...（可能几百张图片，png + captions，甚至有 npz 文件好像是 latent cache??）...</li><li>1boy/ ...</li></ul></li><li>train_plans/<ul><li>reg-my-tran-1/<ul><li>1_1girl/ ...</li><li>1_1boy/ ...</li></ul></li></ul></li></ul></li></ul></li><li>让 AI 写一个脚本，接受一系列参数如 "train:new-test, reg:1girl=200,1boy=100,..." 这样<ul><li>这个脚本，可以自动创建好 reg-new-test，并从指定的 category 里挑选指定数量的图片，创造软链到这个 reg-new-test 文件夹中</li><li>甚至可以追加处理逻辑，当要的 reg 图片多于现有某个 category 的图片，可以自动处理成如下效果：<ul><li>比如 1girl=100，但现在库存只有 24，那么最终得到：</li><li>4_1girl_full/ ...</li><li>1_1girl_rest/ ...（最后4张额外图片）</li></ul></li></ul></li></ul>上面这个过程，会让你感觉更加舒服。<h2 id="how-to-use-a1ahmtbh2">How to Use</h2>In short:<ul><li>Use beige pantyhose<ul><li>If you feel there's no "skin tone," you can add a weaker [[yellow pantyhose]] (weight can be adjusted)</li></ul></li><li>For lora weight, 0.5 ~ 0.6 seems sufficient</li><li>Use 'sheer golden pantyhose', instead of 'golden pantyhose' alone</li></ul><h2 id="optimization-and-findings-naxp68pzz">Optimization and Findings</h2>Below are improvements made during this lora training, which may be helpful for those interested in training lora.<h3 id="1)-(very-important)-image-quality-issues-may-exist-in-details-risbvy6fk">1) (Very Important) Image Quality Issues May Exist in Details</h3>Here are two comparison examples<ul><li>example 1: Carefully look at the overall &amp; left side (problematic) clothing pixel jitter</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/dc03ed7c-f175-4c02-b912-5ee8c1d13de4/width=525/dc03ed7c-f175-4c02-b912-5ee8c1d13de4.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6ba17aba-ec82-4dce-aa2d-9e50dfbc7a00/width=525/6ba17aba-ec82-4dce-aa2d-9e50dfbc7a00.jpeg" /><ul><li>example 2: After enlarging the pantyhose area, the color gradient changes on the left side (problematic)</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8b1926ff-696a-43df-8c03-fd67f04fab74/width=525/8b1926ff-696a-43df-8c03-fd67f04fab74.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d03e8f6e-b87a-4ac8-b29e-3af916bd2bd4/width=525/d03e8f6e-b87a-4ac8-b29e-3af916bd2bd4.jpeg" />As you can see, there's no major problem with the overall image, but because I wanted to improve image quality (as mentioned in 2) below), my previous approach was:<ul><li>First generate the original image</li><li>Then use img2img &amp; adetailer to enlarge it by 1.5 times</li><li>Then use a program to reduce it back to training size &lt;=== This step caused the issue</li></ul>When checking image quality, I didn't zoom in to the pixel level, otherwise I would have seen a lot of noise and color fluctuations.Humans can't perceive this, but the sd model can fully detect it.I also present some comparison experiments from 3 attempts:<ul><li>10_11: fixed dataset &amp; reg set</li><li>10_10: only fixed dataset &amp; wrong reg set</li><li>10_04: wrong dataset &amp; reg set</li></ul>You can see that in 10_04, when the weight increases, the image composition changes dramatically, and this change is meaningless...If you happen to use a program (nodejs sharp dependency) for png processing, you can refer to the following:<pre><code>const sharpInstance = sharp(inputPath)
 .resize(
 width,
 height,
 {kernel: sharp.kernel.lanczos3}
 )
 .png({
 quality: 100,
 // 0~9 is all lossless compression, 9 is better
 compressionLevel: 9,
 // @note. ****** very very very important ******
 palette: false
 });
</code></pre><h3 id="2)-a-technique-for-image-quality-processing-410j2g1nw">2) A Technique for Image Quality Processing</h3>Although 2) is the cause of 1), I still want to recommend it:Initially, to get some images for reg, I directly used training-sized images, generated a batch, and selected from them.But later I found that using highres fix first, or enlarging with img2img, significantly improves the image quality on a macro level, especially when characters are small, as shown below ( left is without upscale and downscale ):<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/bbc83c39-6434-49c6-ab5e-6a500883cc0e/width=525/bbc83c39-6434-49c6-ab5e-6a500883cc0e.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/994a6214-de31-4ad1-8798-378f05016e99/width=525/994a6214-de31-4ad1-8798-378f05016e99.jpeg" />So, this processing approach is:<ul><li>First generate a large batch of images based on prompts, then select a batch that is reasonable and meets expectations</li><li>Enlarge with img2img &amp; adetailer</li><li>Then reduce back to training size</li></ul><h3 id="3)-management-of-reg-sets-tjw2j2l11">3) Management of reg sets</h3>Collecting reg sets takes as much time as datasets, and in a single training session, it requires more quantity and potentially more time.The good news is that most high-quality reg sets can be reused in different trainings 🤩; unfortunately, managing them for such reuse can be challenging 😂.I've verified that:(At least in mac &amp; linux environments) resources like images can be normally used by training scripts when using symbolic links.So I recommend the following approach:<ul><li>Prepare a folder (structure approximately as follows):<ul><li>my_reg_set/<ul><li>all_images/<ul><li>1girl/ ... (possibly hundreds of images, png + captions, even npz files which seem to be latent cache??) ...</li><li>1boy/ ...</li></ul></li><li>train_plans/<ul><li>reg-my-tran-1/<ul><li>1_1girl/ ...</li><li>1_1boy/ ...</li></ul></li></ul></li></ul></li></ul></li><li>Have AI write a script that accepts parameters like "train:new-test, reg:1girl=200,1boy=100,..."<ul><li>This script can automatically create reg-new-test and select the specified number of images from the specified categories, creating symbolic links to this reg-new-test folder</li><li>It can even add processing logic for when the required reg images exceed the available images in a category, resulting in something like:<ul><li>For example, if 1girl=100, but current stock only has 24, then you ultimately get:</li><li>4_1girl_full/ ...</li><li>1_1girl_rest/ ... (the last 4 extra images)</li></ul></li></ul></li></ul>The process above will make you feel more comfortable.<h1 id="previous-doc-uvp1tas2p">==== previous doc ====</h1><h2 id="how-to-use-2dxz8v8k6">How to use</h2><h3 id="-10zctf8lc"></h3><h3 id="what's-the-trigger-tokens-(-for-better-understanding-)-ugzinp20c">What's the trigger tokens ? ( for better understanding )</h3><ul><li>In my dataset (which will be published as well), I use the following trigger tokens:<ul><li>// for controlling the color</li><li>golden pantyhose // more yellow toned<ul><li>Be aware that golden pantyhose is kinda "too strong" in style</li></ul></li><li>beige pantyhose</li><li>// for controlling shininess</li><li>cotton pantyhose</li><li>shiny pantyhose</li></ul></li></ul><h3 id="how-the-actual-result-(-i-might-keep-updating-when-testing-the-lora-...-)-r4ojo1jma">How the actual result? ( I might keep updating when testing the LoRA ... )</h3><ul><li>NOTE : the dataset are mainly depicting "legging", and also tends to "nsfw"<ul><li>so, if you found torn / see-through is not proper, you can specify them explicitly on NEGATIVE PROMPT ( or POSITIVE PROMPT if you do want them )</li></ul></li></ul><ul><li>NOTE : I found the weight 0.6 ~ 0.8 is better, with higher weight, the LoRA will kinda introduce more unrelated details from the dataset (which does not make sense from time to time)</li></ul>v3<ul><li>the closest effect as I expect of nude pantyhose (MAYBE?)</li></ul><pre><code>pos : sheer pantyhose,[yellow pantyhose], beige pantyhose,shiny pantyhose,

neg : torn,[golden pantyhose],</code></pre>v2<pre><code>pos: beige pantyhose, cotton pantyhose,</code></pre>v1<pre><code>pos: golden pantyhose, shiny pantyhose, sheer pantyhose,

neg: (oily pantyhose),</code></pre><h3 id="what-tags-you-can-try-on-pos-neg-prompts-06dmxz22x">What tags you can try on pos / neg prompts ?</h3><pre><code>beige pantyhose,
cotton pantyhose,
shiny pantyhose,
oily pantyhose, // mainly put in neg, otherwise too oily
golden pantyhose, // sometimes you can consider putting it on neg</code></pre>&lt; to be continue &gt;

Lab @10 : nude pantyhose, natural pantyhose, with cotton or shiny controlling

benchmark.zip

[benchmark] 11, 10, 04

Files with SHA256: 560d415c2c63f0f446b43cb759c70984542bd50913e0011d655e0f499246e720